# Cerebras wafer-scale chip

29 Hacker News commenters recommended Cerebras wafer-scale chip and 8 criticized it, November 2025 to November 2026. It was recommended in 10 of 13 months. It meets the Hacker Approved badge rule.

**Verdict:** Wafer-scale chip that people say runs LLM inference far faster than GPUs. Critics agree it is costly and struggles with large models.

## Why people recommend it

- Very fast token throughput (7 comments): "The tokens per second performance numbers coming from Cerebrus/Talas are several orders of magnitude higher than models running on GPUs" ([source](https://news.ycombinator.com/item?id=48981379)) "With the next generation of Cerebras chips expected to be 5–7× faster, peak throughput could reach the ~17,500 tokens-per-second range." ([source](https://news.ycombinator.com/item?id=47097763)) "the advantages of running >5x the token throughput per thread can't be overstated." ([source](https://news.ycombinator.com/item?id=48863811))
- Admired engineering and hardware (7 comments): "Cerebras did the Really Hard Thing, something many other brilliant companies and people failed at for decades." ([source](https://news.ycombinator.com/item?id=47006880)) "takes a lot of cool engineering and cooling to make it work, and is very cool" ([source](https://news.ycombinator.com/item?id=48027137)) "One I like in particular is Cerebras." ([source](https://news.ycombinator.com/item?id=47003171))
- Large on-chip SRAM (2 comments): "Cerebras is kind of insane in that they've built one massive wafer-scale chip with a comparative ocean of SRAM" ([source](https://news.ycombinator.com/item?id=47023043)) "Just doing what Cerebras did and making a wafer-sized chip which is mostly SRAM is a much better solution" ([source](https://news.ycombinator.com/item?id=49207348))

## What people warn about

- High cost and expense (5 comments): "Using a waffer sized chip doesn't sound great from a cost perspective when compared to using many smaller chips for inference." ([source](https://news.ycombinator.com/item?id=46999713)) "makes Cerebras rather cost intensive given that Nvidia can get dozens of GPUs out of a single wafer" ([source](https://news.ycombinator.com/item?id=48981546)) "The initial cost of serving is very high, and while super performant not great for scaling up." ([source](https://news.ycombinator.com/item?id=47812384))
- Struggles with large models (4 comments): "I am skeptical that it will be economically useful for models that are larger in size" ([source](https://news.ycombinator.com/item?id=46999992)) "just need way to many racks to be able to serve the full Deepseek model" ([source](https://news.ycombinator.com/item?id=46629744)) "a multi-million dollar machine with ~40GB of SRAM is not a good solution for big LLMs" ([source](https://news.ycombinator.com/item?id=49487268))
- Little public access (2 comments): "They don't have it, nor do they have a way to get it since they are last in line at the fabs." ([source](https://news.ycombinator.com/item?id=46993881)) "It is not real in the sense that almost nobody has access to them." ([source](https://news.ycombinator.com/item?id=46509058))
- Not practical for scaling (2 comments): "can give great result in the long run, but at the pace AI is growing it's not very practical yet" ([source](https://news.ycombinator.com/item?id=49519401)) "It doesn't scale, and won't ever be profitable." ([source](https://news.ycombinator.com/item?id=49885019))

## Alternatives hackers recommend

- [Google TPU](https://hackerapproved.com/p/google-tpu.md): 73 recommend
- [Taalas HC1](https://hackerapproved.com/p/taalas-chip.md): 59 recommend
- [Taalas hardwired LLM chip](https://hackerapproved.com/p/taalas.md): 45 recommend
- [Cerebras CS-4](https://hackerapproved.com/p/cerebras-cs4.md): 12 recommend
- [Groq LPU](https://hackerapproved.com/p/groq-lpu.md): 9 recommend

## Where to get it

- [Visit the official site](https://www.cerebras.ai/)

## Cite as

Hacker Approved, "Cerebras wafer-scale chip", https://hackerapproved.com/p/cerebras-wafer-scale-chip/ (counts from 1,795,640 Hacker News comments, November 2025 to November 2026).